Bias Detection in AI Models Using Word Embedding Association Tests (WEAT)

Bias in artificial intelligence (AI) systems has emerged as a critical concern, particularly in natural language processing (NLP), where pretrained word embeddings often encode and amplify societal stereotypes. The Word Embedding Association Test (WEAT) has become a foundational method for detecting such biases by quantifying the associative relationships between conceptually relevant word groups. This review provides a comprehensive synthesis of research and applications related to WEAT and its methodological extensions, covering developments. It begins with a technical overview of static and contextual word embeddings and outlines how various forms of bias manifest within them. We explore the origins and mathematical framework of WEAT, followed by its adaptations such as SEAT and ROME that address limitations in modern transformer-based models. The article then examines practical applications of WEAT in sentiment analysis, recommendation systems, healthcare diagnostics, and legal risk assessments, highlighting its role in AI governance and ethical auditing. A comparative analysis of prominent WEAT-supporting tools and empirical case studies further illustrate its effectiveness and limitations. Finally, we discuss unresolved challenges related to model contextuality, definitional ambiguity, and cross-disciplinary integration. Through this review, we underscore the importance of embedding-level bias audits and advocate for the evolution of WEAT into a more context-aware and policy-aligned framework for responsible AI